| Contents | This talk revisits the “3Rs” in computer vision: Recognition, Reconstruction, and Reorganization in the era of modern deep learning. It presents a series of works showing how one R can enhance another. CAST highlights how recognition benefits from and also guides hierarchical segmentation. SHED shows how segmentation grounds dense prediction tasks such as depth estimation, leading to structured 3D reconstruction. OAK illustrates how recognition guides reorganization by contextualizing visual features for open and ad-hoc categorization. Together, these studies demonstrate that the 3Rs provide a structural prior that unifies diverse vision tasks, and invite us to ask: Do today’s foundation models truly capture the 3Rs, and how might we encode them?
|
